A randomized, blinded study presented at the American Heart Association’s Scientific Sessions 2026 found that cardiologists reviewing AI-drafted echocardiogram reports spent an average of 4.6 minutes per exam, compared with 8.5 minutes without AI assistance. The meeting findings also reported similar rates of substantial changes to AI- and human-prepared drafts. The results are preliminary conference evidence—not proof that AI can diagnose independently, improve patient outcomes, or replace a physician’s review.

What the study found

AI-generated preliminary echocardiogram interpretations were associated with less cardiologist report-finalization time in AI ECHO INSIGHT, a study presented November 9 at the American Heart Association (AHA) Scientific Sessions 2026. According to the AHA’s conference news release, the average time was 4.6 minutes per exam when a cardiologist reviewed an AI-generated preliminary report and finalized it, versus 8.5 minutes to complete an interpretation without AI assistance. The release describes that difference as a roughly 46% reduction—close to half.

The headline comparison is not that a machine independently read and signed off on the scans. The workflow retained a cardiologist who reviewed the preliminary interpretation and finalized the report. The stated time measure concerns the cardiologist’s work on the interpretation, not the entire process of acquiring an echocardiogram, delivering care, or reaching a patient outcome. The findings do not establish how much time a hospital would save across all related work or whether a patient would receive a result sooner.

The AHA described the results as a late-breaking science abstract and explicitly cautioned that conference abstracts are not peer-reviewed and should be considered preliminary until a full results manuscript appears in a peer-reviewed journal. The study’s trial protocol, by contrast, has been published in Contemporary Clinical Trials. That paper documents the planned methods; it is not the peer-reviewed final report of these newly presented outcome results. Keeping those two publications distinct matters when assessing what is and is not established.

How the study was designed

The trial protocol describes AI ECHO INSIGHT as a prospective, randomized, blinded clinical study conducted at Kaiser Permanente Northern California. The research evaluated 1,200 historical transthoracic echocardiograms, commonly called TTEs. The protocol says the sample comprised 1,000 consecutive, unselected studies plus 200 studies involving moderate or greater valvular disease. The AHA release says the exams had been acquired during routine care over two weeks in March 2025 at hospital and outpatient settings within 21 medical centers in the Kaiser Permanente Northern California system.

Three preliminary-report workflows were compared: an AI-generated draft, a cardiologist-generated draft, and a sonographer-generated draft. The corresponding reports were assigned to cardiologists for review and finalization. The AHA says reviewers did not know whether a preliminary interpretation had been prepared by the AI system, a sonographer, or another cardiologist. This blinding is important because awareness of the draft’s source could influence how carefully a reader scrutinizes it or how much editing they make.

The AHA release reports participation by nine sonographers, with an average of 7.2 years of experience, and 20 cardiologists, averaging 11.7 years of experience. Each of the three sources prepared preliminary reports for the same 1,200 echocardiograms, and cardiologists finalized the assigned drafts. The release characterizes the experiment as randomized; the protocol gives the overall framework and groups the human-generated reports for the principal comparison. These are details of a structured research workflow, not evidence that the AI was deployed routinely for all patients at those centers.

The protocol paper, published in October 2026 with online publication in August, is useful for understanding the prespecified design. It identifies the primary outcome as the rate of substantial changes between preliminary and final reports, comparing the AI-assisted workflow with pooled cardiologist-assisted and sonographer-assisted workflows. Planned secondary outcomes included interpretation time, diagnostic-accuracy analyses, and reporting consistency. The protocol establishes what investigators intended to measure; the conference report supplies the results currently being discussed.

What the time result means—and does not mean

The AHA reported 4.6 minutes for cardiologists reviewing and finalizing AI-generated drafts, compared with 8.5 minutes to generate reports without AI assistance. The reported difference is 3.9 minutes per exam. The release summarizes it as a 46% time saving. Those averages offer a potentially meaningful workflow signal: if replicated under ordinary conditions, reducing report-finalization time could let a practice use clinicians’ time differently. They do not, by themselves, show that every individual case was faster or that the same reduction would hold in every department.

Interpretation time is only one part of clinical workload. A report may require additional image review, comparison with prior studies, communication with the ordering team, or reconciliation with a patient’s history. The available release does not provide enough detail to quantify all such tasks or to determine whether the time comparison includes every component of the clinical process. Nor does the time result demonstrate improved diagnostic accuracy. The accuracy-related endpoint and the efficiency endpoint answer separate questions.

It is also important to avoid converting an average research finding into a guaranteed operational saving. The study used prior routine-care exams but asked clinicians to work in a research setting. The AHA notes that participating cardiologists knew they were in a study and may have changed their usual work patterns. Performance can also depend on local reporting systems, case mix, staffing, the AI tool’s integration, and how much editing a particular examination requires. Any implementation would need to measure local outcomes rather than assume the reported average transfers directly.

How the comparison endpoint worked

“Comparable” in this report refers to the study’s chosen measure: how often a cardiologist made a substantial change to the preliminary interpretation before finalizing it. The AHA reported substantial changes to 50.6% of AI-generated preliminary interpretations and 49.7% of preliminary interpretations prepared by humans. The researchers said the comparison met the trial’s non-inferiority criteria, meaning the AI-assisted workflow met a prespecified standard for being no worse than the human-assisted comparison on that endpoint.

This is not the same as saying that the AI was correct in 50.6% of cases, or that the final reports were error-free. A “substantial change” is an editing-based proxy, not a direct tally of clinically consequential diagnostic errors against an independent gold standard. A draft can need correction without a substantial edit, and a substantial edit does not automatically mean the initial report would have caused harm. The endpoint helps compare workflow acceptability under the trial’s definitions; it should not be mistaken for a universal accuracy rate.

The AHA additionally said AI-assisted reports were more consistent with historical clinical reports in the medical record, and that reviewers could not identify whether the draft came from AI or a human. Those findings offer context, but the release does not make historical reports an independent ground truth. Consistency with prior documentation may be useful, but it cannot on its own establish that a report is clinically correct. The abstract’s conference status and the limits of the chosen endpoints remain relevant when interpreting these secondary observations.

Limitations and unanswered questions

The investigators did not test whether AI assistance improved long-term health outcomes for patients. That limitation is explicit in the AHA’s release. The study therefore cannot establish that the tool reduces complications, leads to more appropriate treatment, prevents missed disease, shortens time to treatment, or improves survival. A faster report workflow may be valuable, but patient benefit requires evidence connecting the workflow change to clinically meaningful outcomes.

Generalizability is another open question. The exams came from one integrated health system in Northern California, even though they were collected across 21 medical centers and included hospital and outpatient settings. The AHA reports a patient average age of 69; 47% were women and 53% men. The release provides race and ethnicity categories for the study population, but a sample from one regional system cannot represent every healthcare setting, patient population, imaging device, or reporting practice. Larger-scale evaluation across different systems is needed.

The exam set included selected cases with moderate or greater valvular disease alongside consecutive unselected exams. That design may help examine a broader range of abnormalities, but the results cannot automatically be generalized to every type of echocardiogram, rare finding, image-quality challenge, or patient subgroup. The publicly available conference release does not give all subgroup estimates, the full statistical analysis, or enough information to independently assess every performance measure. Those details should be checked against the complete results paper if and when it is published.

Finally, study participation itself can alter behavior. The AHA says clinicians knew they were being observed and may have adjusted their usual work patterns. A controlled trial can isolate a comparison more carefully than routine practice, but routine deployment introduces interruptions, software integration issues, accountability questions, and variation in how clinicians use suggestions. As the investigators indicated, further work should assess performance at larger scale and in real-world clinical settings.

Clinical context: assistance, not autonomous diagnosis

An echocardiogram is an ultrasound examination of the heart that clinicians use to assess cardiac structures and function. The workflow described by the AHA has a sonographer acquire images and a cardiologist interpret those images and produce a report. In the tested AI-assisted pathway, software generates a preliminary interpretation; a cardiologist then reviews and finalizes it. The distinction between a draft and a signed clinical interpretation is central to understanding the result.

Co-senior author Keane K. Lee told the AHA that the cardiologist remains ultimately responsible for interpreting the echocardiogram and certifying the final report, including when AI assists. The clinician also brings patient-specific context and judgment to the final interpretation and care plan. The AHA release does not announce regulatory clearance or approval for autonomous diagnosis, and these research findings should not be described as such. Nothing in the presented study shows the software replacing the physician or making a final diagnosis without oversight.

In practical terms, the research asks whether an AI-generated preliminary report can support a human reader’s workflow while meeting a prespecified comparison criterion. It does not answer whether a hospital should adopt a particular tool, how responsibility should be allocated, or what safeguards are needed in a specific deployment. Those decisions require evidence about the specific product and setting, appropriate governance, monitoring for errors, and clear accountability for final reports.

What to watch next

The most consequential next step is a full manuscript reporting the trial’s results, statistical methods, confidence intervals, subgroup analyses, and any implementation details that the conference announcement does not include. Because the AHA calls the findings preliminary, readers should not treat the release alone as a substitute for a peer-reviewed results paper. The already published protocol helps readers see the intended design, but a protocol cannot validate the observed time or editing outcomes.

Further studies should test whether the reported time difference persists in ordinary clinical practice and whether it changes the time from image acquisition to a signed report. They should examine diagnostic performance using carefully defined reference standards, quantify clinically meaningful errors, and assess effects across different populations, sites, devices, and levels of image quality. Patient outcomes, downstream testing, and the possibility of automation bias—when a clinician is overly influenced by a software suggestion—also warrant evaluation.

For now, the measured takeaway is narrow but potentially useful: in a blinded randomized research workflow involving 1,200 exams, AHA-reported substantial-change rates were similar for AI and human preliminary reports, and average interpretation time was 4.6 rather than 8.5 minutes with AI drafts. The results support continued evaluation of AI as a clinician-assistance tool. They do not establish autonomous diagnosis, regulatory status, improved patient outcomes, or guaranteed time savings beyond the studied setting.

Frequently asked questions

How many echocardiograms were studied?

The AHA release reports 1,200 historical echocardiograms from routine care across 21 Kaiser Permanente Northern California medical centers. The published protocol describes 1,000 consecutive, unselected studies and 200 with moderate or greater valvular disease.

How many clinicians took part?

The AHA release reports nine sonographers and 20 cardiologists. The sonographers averaged 7.2 years of experience, while cardiologists averaged 11.7 years.

What did 50.6% versus 49.7% measure?

It was the proportion of preliminary interpretations that cardiologists substantially changed before finalizing them: 50.6% for AI-generated drafts versus 49.7% for human-generated drafts, according to the AHA. It is not an overall diagnostic accuracy percentage.

Was AI faster in the study?

The AHA reported average interpretation times of 4.6 minutes with an AI-generated preliminary report and 8.5 minutes without AI assistance, describing the difference as about 46%. It is a study average, not a guarantee for each scan or clinical site.

Are the findings peer-reviewed or evidence of autonomous diagnosis?

The outcome findings were presented as an AHA conference abstract, which the AHA says is not peer-reviewed and is preliminary until published as a full results manuscript. The trial tested AI as a drafting aid with cardiologist review and finalization; it does not establish autonomous diagnosis or regulatory approval.

References